net/sunrpc/xprt.c

   1 /*
   2  *  linux/net/sunrpc/xprt.c
   3  *
   4  *  This is a generic RPC call interface supporting congestion avoidance,
   5  *  and asynchronous calls.
   6  *
   7  *  The interface works like this:
   8  *
   9  *  -   When a process places a call, it allocates a request slot if
  10  *      one is available. Otherwise, it sleeps on the backlog queue
  11  *      (xprt_reserve).
  12  *  -   Next, the caller puts together the RPC message, stuffs it into
  13  *      the request struct, and calls xprt_transmit().
  14  *  -   xprt_transmit sends the message and installs the caller on the
  15  *      transport's wait list. At the same time, if a reply is expected,
  16  *      it installs a timer that is run after the packet's timeout has
  17  *      expired.
  18  *  -   When a packet arrives, the data_ready handler walks the list of
  19  *      pending requests for that transport. If a matching XID is found, the
  20  *      caller is woken up, and the timer removed.
  21  *  -   When no reply arrives within the timeout interval, the timer is
  22  *      fired by the kernel and runs xprt_timer(). It either adjusts the
  23  *      timeout values (minor timeout) or wakes up the caller with a status
  24  *      of -ETIMEDOUT.
  25  *  -   When the caller receives a notification from RPC that a reply arrived,
  26  *      it should release the RPC slot, and process the reply.
  27  *      If the call timed out, it may choose to retry the operation by
  28  *      adjusting the initial timeout value, and simply calling rpc_call
  29  *      again.
  30  *
  31  *  Support for async RPC is done through a set of RPC-specific scheduling
  32  *  primitives that `transparently' work for processes as well as async
  33  *  tasks that rely on callbacks.
  34  *
  35  *  Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
  36  *
  37  *  Transport switch API copyright (C) 2005, Chuck Lever <cel@netapp.com>
  38  */
  39
  40 #include <linux/module.h>
  41
  42 #include <linux/types.h>
  43 #include <linux/interrupt.h>
  44 #include <linux/workqueue.h>
  45 #include <linux/net.h>
  46 #include <linux/ktime.h>
  47
  48 #include <linux/sunrpc/clnt.h>
  49 #include <linux/sunrpc/metrics.h>
  50 #include <linux/sunrpc/bc_xprt.h>
  51
  52 #include "sunrpc.h"
  53
  54 /*
  55  * Local variables
  56  */
  57
  58 #ifdef RPC_DEBUG
  59 # define RPCDBG_FACILITY        RPCDBG_XPRT
  60 #endif
  61
  62 /*
  63  * Local functions
  64  */
  65 static void      xprt_init(struct rpc_xprt *xprt, struct net *net);
  66 static void     xprt_request_init(struct rpc_task *, struct rpc_xprt *);
  67 static void     xprt_connect_status(struct rpc_task *task);
  68 static int      __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
  69 static void      xprt_destroy(struct rpc_xprt *xprt);
  70
  71 static DEFINE_SPINLOCK(xprt_list_lock);
  72 static LIST_HEAD(xprt_list);
  73
  74 /*
  75  * The transport code maintains an estimate on the maximum number of out-
  76  * standing RPC requests, using a smoothed version of the congestion
  77  * avoidance implemented in 44BSD. This is basically the Van Jacobson
  78  * congestion algorithm: If a retransmit occurs, the congestion window is
  79  * halved; otherwise, it is incremented by 1/cwnd when
  80  *
  81  *      -       a reply is received and
  82  *      -       a full number of requests are outstanding and
  83  *      -       the congestion window hasn't been updated recently.
  84  */
  85 #define RPC_CWNDSHIFT           (8U)
  86 #define RPC_CWNDSCALE           (1U << RPC_CWNDSHIFT)
  87 #define RPC_INITCWND            RPC_CWNDSCALE
  88 #define RPC_MAXCWND(xprt)       ((xprt)->max_reqs << RPC_CWNDSHIFT)
  89
  90 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
  91
  92 /**
  93  * xprt_register_transport - register a transport implementation
  94  * @transport: transport to register
  95  *
  96  * If a transport implementation is loaded as a kernel module, it can
  97  * call this interface to make itself known to the RPC client.
  98  *
  99  * Returns:
 100  * 0:           transport successfully registered
 101  * -EEXIST:     transport already registered
 102  * -EINVAL:     transport module being unloaded
 103  */
 104 int xprt_register_transport(struct xprt_class *transport)
 105 {
 106         struct xprt_class *t;
 107         int result;
 108
 109         result = -EEXIST;
 110         spin_lock(&xprt_list_lock);
 111         list_for_each_entry(t, &xprt_list, list) {
 112                 /* don't register the same transport class twice */
 113                 if (t->ident == transport->ident)
 114                         goto out;
 115         }
 116
 117         list_add_tail(&transport->list, &xprt_list);
 118         printk(KERN_INFO "RPC: Registered %s transport module.\n",
 119                transport->name);
 120         result = 0;
 121
 122 out:
 123         spin_unlock(&xprt_list_lock);
 124         return result;
 125 }
 126 EXPORT_SYMBOL_GPL(xprt_register_transport);
 127
 128 /**
 129  * xprt_unregister_transport - unregister a transport implementation
 130  * @transport: transport to unregister
 131  *
 132  * Returns:
 133  * 0:           transport successfully unregistered
 134  * -ENOENT:     transport never registered
 135  */
 136 int xprt_unregister_transport(struct xprt_class *transport)
 137 {
 138         struct xprt_class *t;
 139         int result;
 140
 141         result = 0;
 142         spin_lock(&xprt_list_lock);
 143         list_for_each_entry(t, &xprt_list, list) {
 144                 if (t == transport) {
 145                         printk(KERN_INFO
 146                                 "RPC: Unregistered %s transport module.\n",
 147                                 transport->name);
 148                         list_del_init(&transport->list);
 149                         goto out;
 150                 }
 151         }
 152         result = -ENOENT;
 153
 154 out:
 155         spin_unlock(&xprt_list_lock);
 156         return result;
 157 }
 158 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
 159
 160 /**
 161  * xprt_load_transport - load a transport implementation
 162  * @transport_name: transport to load
 163  *
 164  * Returns:
 165  * 0:           transport successfully loaded
 166  * -ENOENT:     transport module not available
 167  */
 168 int xprt_load_transport(const char *transport_name)
 169 {
 170         struct xprt_class *t;
 171         int result;
 172
 173         result = 0;
 174         spin_lock(&xprt_list_lock);
 175         list_for_each_entry(t, &xprt_list, list) {
 176                 if (strcmp(t->name, transport_name) == 0) {
 177                         spin_unlock(&xprt_list_lock);
 178                         goto out;
 179                 }
 180         }
 181         spin_unlock(&xprt_list_lock);
 182         result = request_module("xprt%s", transport_name);
 183 out:
 184         return result;
 185 }
 186 EXPORT_SYMBOL_GPL(xprt_load_transport);
 187
 188 /**
 189  * xprt_reserve_xprt - serialize write access to transports
 190  * @task: task that is requesting access to the transport
 191  * @xprt: pointer to the target transport
 192  *
 193  * This prevents mixing the payload of separate requests, and prevents
 194  * transport connects from colliding with writes.  No congestion control
 195  * is provided.
 196  */
 197 int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 198 {
 199         struct rpc_rqst *req = task->tk_rqstp;
 200         int priority;
 201
 202         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 203                 if (task == xprt->snd_task)
 204                         return 1;
 205                 goto out_sleep;
 206         }
 207         xprt->snd_task = task;
 208         if (req != NULL) {
 209                 req->rq_bytes_sent = 0;
 210                 req->rq_ntrans++;
 211         }
 212
 213         return 1;
 214
 215 out_sleep:
 216         dprintk("RPC: %5u failed to lock transport %p\n",
 217                         task->tk_pid, xprt);
 218         task->tk_timeout = 0;
 219         task->tk_status = -EAGAIN;
 220         if (req == NULL)
 221                 priority = RPC_PRIORITY_LOW;
 222         else if (!req->rq_ntrans)
 223                 priority = RPC_PRIORITY_NORMAL;
 224         else
 225                 priority = RPC_PRIORITY_HIGH;
 226         rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
 227         return 0;
 228 }
 229 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
 230
 231 static void xprt_clear_locked(struct rpc_xprt *xprt)
 232 {
 233         xprt->snd_task = NULL;
 234         if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state)) {
 235                 smp_mb__before_clear_bit();
 236                 clear_bit(XPRT_LOCKED, &xprt->state);
 237                 smp_mb__after_clear_bit();
 238         } else
 239                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 240 }
 241
 242 /*
 243  * xprt_reserve_xprt_cong - serialize write access to transports
 244  * @task: task that is requesting access to the transport
 245  *
 246  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
 247  * integrated into the decision of whether a request is allowed to be
 248  * woken up and given access to the transport.
 249  */
 250 int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 251 {
 252         struct rpc_rqst *req = task->tk_rqstp;
 253         int priority;
 254
 255         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 256                 if (task == xprt->snd_task)
 257                         return 1;
 258                 goto out_sleep;
 259         }
 260         if (req == NULL) {
 261                 xprt->snd_task = task;
 262                 return 1;
 263         }
 264         if (__xprt_get_cong(xprt, task)) {
 265                 xprt->snd_task = task;
 266                 req->rq_bytes_sent = 0;
 267                 req->rq_ntrans++;
 268                 return 1;
 269         }
 270         xprt_clear_locked(xprt);
 271 out_sleep:
 272         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
 273         task->tk_timeout = 0;
 274         task->tk_status = -EAGAIN;
 275         if (req == NULL)
 276                 priority = RPC_PRIORITY_LOW;
 277         else if (!req->rq_ntrans)
 278                 priority = RPC_PRIORITY_NORMAL;
 279         else
 280                 priority = RPC_PRIORITY_HIGH;
 281         rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
 282         return 0;
 283 }
 284 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
 285
 286 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 287 {
 288         int retval;
 289
 290         spin_lock_bh(&xprt->transport_lock);
 291         retval = xprt->ops->reserve_xprt(xprt, task);
 292         spin_unlock_bh(&xprt->transport_lock);
 293         return retval;
 294 }
 295
 296 static bool __xprt_lock_write_func(struct rpc_task *task, void *data)
 297 {
 298         struct rpc_xprt *xprt = data;
 299         struct rpc_rqst *req;
 300
 301         req = task->tk_rqstp;
 302         xprt->snd_task = task;
 303         if (req) {
 304                 req->rq_bytes_sent = 0;
 305                 req->rq_ntrans++;
 306         }
 307         return true;
 308 }
 309
 310 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
 311 {
 312         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 313                 return;
 314
 315         if (rpc_wake_up_first(&xprt->sending, __xprt_lock_write_func, xprt))
 316                 return;
 317         xprt_clear_locked(xprt);
 318 }
 319
 320 static bool __xprt_lock_write_cong_func(struct rpc_task *task, void *data)
 321 {
 322         struct rpc_xprt *xprt = data;
 323         struct rpc_rqst *req;
 324
 325         req = task->tk_rqstp;
 326         if (req == NULL) {
 327                 xprt->snd_task = task;
 328                 return true;
 329         }
 330         if (__xprt_get_cong(xprt, task)) {
 331                 xprt->snd_task = task;
 332                 req->rq_bytes_sent = 0;
 333                 req->rq_ntrans++;
 334                 return true;
 335         }
 336         return false;
 337 }
 338
 339 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
 340 {
 341         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 342                 return;
 343         if (RPCXPRT_CONGESTED(xprt))
 344                 goto out_unlock;
 345         if (rpc_wake_up_first(&xprt->sending, __xprt_lock_write_cong_func, xprt))
 346                 return;
 347 out_unlock:
 348         xprt_clear_locked(xprt);
 349 }
 350
 351 /**
 352  * xprt_release_xprt - allow other requests to use a transport
 353  * @xprt: transport with other tasks potentially waiting
 354  * @task: task that is releasing access to the transport
 355  *
 356  * Note that "task" can be NULL.  No congestion control is provided.
 357  */
 358 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 359 {
 360         if (xprt->snd_task == task) {
 361                 if (task != NULL) {
 362                         struct rpc_rqst *req = task->tk_rqstp;
 363                         if (req != NULL)
 364                                 req->rq_bytes_sent = 0;
 365                 }
 366                 xprt_clear_locked(xprt);
 367                 __xprt_lock_write_next(xprt);
 368         }
 369 }
 370 EXPORT_SYMBOL_GPL(xprt_release_xprt);
 371
 372 /**
 373  * xprt_release_xprt_cong - allow other requests to use a transport
 374  * @xprt: transport with other tasks potentially waiting
 375  * @task: task that is releasing access to the transport
 376  *
 377  * Note that "task" can be NULL.  Another task is awoken to use the
 378  * transport if the transport's congestion window allows it.
 379  */
 380 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 381 {
 382         if (xprt->snd_task == task) {
 383                 if (task != NULL) {
 384                         struct rpc_rqst *req = task->tk_rqstp;
 385                         if (req != NULL)
 386                                 req->rq_bytes_sent = 0;
 387                 }
 388                 xprt_clear_locked(xprt);
 389                 __xprt_lock_write_next_cong(xprt);
 390         }
 391 }
 392 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
 393
 394 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 395 {
 396         spin_lock_bh(&xprt->transport_lock);
 397         xprt->ops->release_xprt(xprt, task);
 398         spin_unlock_bh(&xprt->transport_lock);
 399 }
 400
 401 /*
 402  * Van Jacobson congestion avoidance. Check if the congestion window
 403  * overflowed. Put the task to sleep if this is the case.
 404  */
 405 static int
 406 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 407 {
 408         struct rpc_rqst *req = task->tk_rqstp;
 409
 410         if (req->rq_cong)
 411                 return 1;
 412         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
 413                         task->tk_pid, xprt->cong, xprt->cwnd);
 414         if (RPCXPRT_CONGESTED(xprt))
 415                 return 0;
 416         req->rq_cong = 1;
 417         xprt->cong += RPC_CWNDSCALE;
 418         return 1;
 419 }
 420
 421 /*
 422  * Adjust the congestion window, and wake up the next task
 423  * that has been sleeping due to congestion
 424  */
 425 static void
 426 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 427 {
 428         if (!req->rq_cong)
 429                 return;
 430         req->rq_cong = 0;
 431         xprt->cong -= RPC_CWNDSCALE;
 432         __xprt_lock_write_next_cong(xprt);
 433 }
 434
 435 /**
 436  * xprt_release_rqst_cong - housekeeping when request is complete
 437  * @task: RPC request that recently completed
 438  *
 439  * Useful for transports that require congestion control.
 440  */
 441 void xprt_release_rqst_cong(struct rpc_task *task)
 442 {
 443         struct rpc_rqst *req = task->tk_rqstp;
 444
 445         __xprt_put_cong(req->rq_xprt, req);
 446 }
 447 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
 448
 449 /**
 450  * xprt_adjust_cwnd - adjust transport congestion window
 451  * @xprt: pointer to xprt
 452  * @task: recently completed RPC request used to adjust window
 453  * @result: result code of completed RPC request
 454  *
 455  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 456  */
 457 void xprt_adjust_cwnd(struct rpc_xprt *xprt, struct rpc_task *task, int result)
 458 {
 459         struct rpc_rqst *req = task->tk_rqstp;
 460         unsigned long cwnd = xprt->cwnd;
 461
 462         if (result >= 0 && cwnd <= xprt->cong) {
 463                 /* The (cwnd >> 1) term makes sure
 464                  * the result gets rounded properly. */
 465                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 466                 if (cwnd > RPC_MAXCWND(xprt))
 467                         cwnd = RPC_MAXCWND(xprt);
 468                 __xprt_lock_write_next_cong(xprt);
 469         } else if (result == -ETIMEDOUT) {
 470                 cwnd >>= 1;
 471                 if (cwnd < RPC_CWNDSCALE)
 472                         cwnd = RPC_CWNDSCALE;
 473         }
 474         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
 475                         xprt->cong, xprt->cwnd, cwnd);
 476         xprt->cwnd = cwnd;
 477         __xprt_put_cong(xprt, req);
 478 }
 479 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
 480
 481 /**
 482  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
 483  * @xprt: transport with waiting tasks
 484  * @status: result code to plant in each task before waking it
 485  *
 486  */
 487 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
 488 {
 489         if (status < 0)
 490                 rpc_wake_up_status(&xprt->pending, status);
 491         else
 492                 rpc_wake_up(&xprt->pending);
 493 }
 494 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
 495
 496 /**
 497  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
 498  * @task: task to be put to sleep
 499  * @action: function pointer to be executed after wait
 500  *
 501  * Note that we only set the timer for the case of RPC_IS_SOFT(), since
 502  * we don't in general want to force a socket disconnection due to
 503  * an incomplete RPC call transmission.
 504  */
 505 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
 506 {
 507         struct rpc_rqst *req = task->tk_rqstp;
 508         struct rpc_xprt *xprt = req->rq_xprt;
 509
 510         task->tk_timeout = RPC_IS_SOFT(task) ? req->rq_timeout : 0;
 511         rpc_sleep_on(&xprt->pending, task, action);
 512 }
 513 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
 514
 515 /**
 516  * xprt_write_space - wake the task waiting for transport output buffer space
 517  * @xprt: transport with waiting tasks
 518  *
 519  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
 520  */
 521 void xprt_write_space(struct rpc_xprt *xprt)
 522 {
 523         spin_lock_bh(&xprt->transport_lock);
 524         if (xprt->snd_task) {
 525                 dprintk("RPC:       write space: waking waiting task on "
 526                                 "xprt %p\n", xprt);
 527                 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
 528         }
 529         spin_unlock_bh(&xprt->transport_lock);
 530 }
 531 EXPORT_SYMBOL_GPL(xprt_write_space);
 532
 533 /**
 534  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 535  * @task: task whose timeout is to be set
 536  *
 537  * Set a request's retransmit timeout based on the transport's
 538  * default timeout parameters.  Used by transports that don't adjust
 539  * the retransmit timeout based on round-trip time estimation.
 540  */
 541 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 542 {
 543         task->tk_timeout = task->tk_rqstp->rq_timeout;
 544 }
 545 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
 546
 547 /**
 548  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 549  * @task: task whose timeout is to be set
 550  *
 551  * Set a request's retransmit timeout using the RTT estimator.
 552  */
 553 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 554 {
 555         int timer = task->tk_msg.rpc_proc->p_timer;
 556         struct rpc_clnt *clnt = task->tk_client;
 557         struct rpc_rtt *rtt = clnt->cl_rtt;
 558         struct rpc_rqst *req = task->tk_rqstp;
 559         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
 560
 561         task->tk_timeout = rpc_calc_rto(rtt, timer);
 562         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 563         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 564                 task->tk_timeout = max_timeout;
 565 }
 566 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
 567
 568 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 569 {
 570         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 571
 572         req->rq_majortimeo = req->rq_timeout;
 573         if (to->to_exponential)
 574                 req->rq_majortimeo <<= to->to_retries;
 575         else
 576                 req->rq_majortimeo += to->to_increment * to->to_retries;
 577         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 578                 req->rq_majortimeo = to->to_maxval;
 579         req->rq_majortimeo += jiffies;
 580 }
 581
 582 /**
 583  * xprt_adjust_timeout - adjust timeout values for next retransmit
 584  * @req: RPC request containing parameters to use for the adjustment
 585  *
 586  */
 587 int xprt_adjust_timeout(struct rpc_rqst *req)
 588 {
 589         struct rpc_xprt *xprt = req->rq_xprt;
 590         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 591         int status = 0;
 592
 593         if (time_before(jiffies, req->rq_majortimeo)) {
 594                 if (to->to_exponential)
 595                         req->rq_timeout <<= 1;
 596                 else
 597                         req->rq_timeout += to->to_increment;
 598                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 599                         req->rq_timeout = to->to_maxval;
 600                 req->rq_retries++;
 601         } else {
 602                 req->rq_timeout = to->to_initval;
 603                 req->rq_retries = 0;
 604                 xprt_reset_majortimeo(req);
 605                 /* Reset the RTT counters == "slow start" */
 606                 spin_lock_bh(&xprt->transport_lock);
 607                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 608                 spin_unlock_bh(&xprt->transport_lock);
 609                 status = -ETIMEDOUT;
 610         }
 611
 612         if (req->rq_timeout == 0) {
 613                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 614                 req->rq_timeout = 5 * HZ;
 615         }
 616         return status;
 617 }
 618
 619 static void xprt_autoclose(struct work_struct *work)
 620 {
 621         struct rpc_xprt *xprt =
 622                 container_of(work, struct rpc_xprt, task_cleanup);
 623
 624         xprt->ops->close(xprt);
 625         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 626         xprt_release_write(xprt, NULL);
 627 }
 628
 629 /**
 630  * xprt_disconnect_done - mark a transport as disconnected
 631  * @xprt: transport to flag for disconnect
 632  *
 633  */
 634 void xprt_disconnect_done(struct rpc_xprt *xprt)
 635 {
 636         dprintk("RPC:       disconnected transport %p\n", xprt);
 637         spin_lock_bh(&xprt->transport_lock);
 638         xprt_clear_connected(xprt);
 639         xprt_wake_pending_tasks(xprt, -EAGAIN);
 640         spin_unlock_bh(&xprt->transport_lock);
 641 }
 642 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
 643
 644 /**
 645  * xprt_force_disconnect - force a transport to disconnect
 646  * @xprt: transport to disconnect
 647  *
 648  */
 649 void xprt_force_disconnect(struct rpc_xprt *xprt)
 650 {
 651         /* Don't race with the test_bit() in xprt_clear_locked() */
 652         spin_lock_bh(&xprt->transport_lock);
 653         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 654         /* Try to schedule an autoclose RPC call */
 655         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 656                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 657         xprt_wake_pending_tasks(xprt, -EAGAIN);
 658         spin_unlock_bh(&xprt->transport_lock);
 659 }
 660
 661 /**
 662  * xprt_conditional_disconnect - force a transport to disconnect
 663  * @xprt: transport to disconnect
 664  * @cookie: 'connection cookie'
 665  *
 666  * This attempts to break the connection if and only if 'cookie' matches
 667  * the current transport 'connection cookie'. It ensures that we don't
 668  * try to break the connection more than once when we need to retransmit
 669  * a batch of RPC requests.
 670  *
 671  */
 672 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
 673 {
 674         /* Don't race with the test_bit() in xprt_clear_locked() */
 675         spin_lock_bh(&xprt->transport_lock);
 676         if (cookie != xprt->connect_cookie)
 677                 goto out;
 678         if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt))
 679                 goto out;
 680         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 681         /* Try to schedule an autoclose RPC call */
 682         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 683                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 684         xprt_wake_pending_tasks(xprt, -EAGAIN);
 685 out:
 686         spin_unlock_bh(&xprt->transport_lock);
 687 }
 688
 689 static void
 690 xprt_init_autodisconnect(unsigned long data)
 691 {
 692         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 693
 694         spin_lock(&xprt->transport_lock);
 695         if (!list_empty(&xprt->recv))
 696                 goto out_abort;
 697         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 698                 goto out_abort;
 699         spin_unlock(&xprt->transport_lock);
 700         set_bit(XPRT_CONNECTION_CLOSE, &xprt->state);
 701         queue_work(rpciod_workqueue, &xprt->task_cleanup);
 702         return;
 703 out_abort:
 704         spin_unlock(&xprt->transport_lock);
 705 }
 706
 707 /**
 708  * xprt_connect - schedule a transport connect operation
 709  * @task: RPC task that is requesting the connect
 710  *
 711  */
 712 void xprt_connect(struct rpc_task *task)
 713 {
 714         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 715
 716         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
 717                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 718
 719         if (!xprt_bound(xprt)) {
 720                 task->tk_status = -EAGAIN;
 721                 return;
 722         }
 723         if (!xprt_lock_write(xprt, task))
 724                 return;
 725
 726         if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
 727                 xprt->ops->close(xprt);
 728
 729         if (xprt_connected(xprt))
 730                 xprt_release_write(xprt, task);
 731         else {
 732                 task->tk_rqstp->rq_bytes_sent = 0;
 733                 task->tk_timeout = task->tk_rqstp->rq_timeout;
 734                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
 735
 736                 if (test_bit(XPRT_CLOSING, &xprt->state))
 737                         return;
 738                 if (xprt_test_and_set_connecting(xprt))
 739                         return;
 740                 xprt->stat.connect_start = jiffies;
 741                 xprt->ops->connect(xprt, task);
 742         }
 743 }
 744
 745 static void xprt_connect_status(struct rpc_task *task)
 746 {
 747         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 748
 749         if (task->tk_status == 0) {
 750                 xprt->stat.connect_count++;
 751                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
 752                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
 753                                 task->tk_pid);
 754                 return;
 755         }
 756
 757         switch (task->tk_status) {
 758         case -EAGAIN:
 759                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
 760                 break;
 761         case -ETIMEDOUT:
 762                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
 763                                 "out\n", task->tk_pid);
 764                 break;
 765         default:
 766                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
 767                                 "server %s\n", task->tk_pid, -task->tk_status,
 768                                 xprt->servername);
 769                 xprt_release_write(xprt, task);
 770                 task->tk_status = -EIO;
 771         }
 772 }
 773
 774 /**
 775  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 776  * @xprt: transport on which the original request was transmitted
 777  * @xid: RPC XID of incoming reply
 778  *
 779  */
 780 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
 781 {
 782         struct rpc_rqst *entry;
 783
 784         list_for_each_entry(entry, &xprt->recv, rq_list)
 785                 if (entry->rq_xid == xid)
 786                         return entry;
 787
 788         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
 789                         ntohl(xid));
 790         xprt->stat.bad_xids++;
 791         return NULL;
 792 }
 793 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
 794
 795 static void xprt_update_rtt(struct rpc_task *task)
 796 {
 797         struct rpc_rqst *req = task->tk_rqstp;
 798         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 799         unsigned int timer = task->tk_msg.rpc_proc->p_timer;
 800         long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
 801
 802         if (timer) {
 803                 if (req->rq_ntrans == 1)
 804                         rpc_update_rtt(rtt, timer, m);
 805                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
 806         }
 807 }
 808
 809 /**
 810  * xprt_complete_rqst - called when reply processing is complete
 811  * @task: RPC request that recently completed
 812  * @copied: actual number of bytes received from the transport
 813  *
 814  * Caller holds transport lock.
 815  */
 816 void xprt_complete_rqst(struct rpc_task *task, int copied)
 817 {
 818         struct rpc_rqst *req = task->tk_rqstp;
 819         struct rpc_xprt *xprt = req->rq_xprt;
 820
 821         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
 822                         task->tk_pid, ntohl(req->rq_xid), copied);
 823
 824         xprt->stat.recvs++;
 825         req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
 826         if (xprt->ops->timer != NULL)
 827                 xprt_update_rtt(task);
 828
 829         list_del_init(&req->rq_list);
 830         req->rq_private_buf.len = copied;
 831         /* Ensure all writes are done before we update */
 832         /* req->rq_reply_bytes_recvd */
 833         smp_wmb();
 834         req->rq_reply_bytes_recvd = copied;
 835         rpc_wake_up_queued_task(&xprt->pending, task);
 836 }
 837 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
 838
 839 static void xprt_timer(struct rpc_task *task)
 840 {
 841         struct rpc_rqst *req = task->tk_rqstp;
 842         struct rpc_xprt *xprt = req->rq_xprt;
 843
 844         if (task->tk_status != -ETIMEDOUT)
 845                 return;
 846         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
 847
 848         spin_lock_bh(&xprt->transport_lock);
 849         if (!req->rq_reply_bytes_recvd) {
 850                 if (xprt->ops->timer)
 851                         xprt->ops->timer(xprt, task);
 852         } else
 853                 task->tk_status = 0;
 854         spin_unlock_bh(&xprt->transport_lock);
 855 }
 856
 857 static inline int xprt_has_timer(struct rpc_xprt *xprt)
 858 {
 859         return xprt->idle_timeout != 0;
 860 }
 861
 862 /**
 863  * xprt_prepare_transmit - reserve the transport before sending a request
 864  * @task: RPC task about to send a request
 865  *
 866  */
 867 bool xprt_prepare_transmit(struct rpc_task *task)
 868 {
 869         struct rpc_rqst *req = task->tk_rqstp;
 870         struct rpc_xprt *xprt = req->rq_xprt;
 871         bool ret = false;
 872
 873         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
 874
 875         spin_lock_bh(&xprt->transport_lock);
 876         if (!req->rq_bytes_sent) {
 877                 if (req->rq_reply_bytes_recvd) {
 878                         task->tk_status = req->rq_reply_bytes_recvd;
 879                         goto out_unlock;
 880                 }
 881                 if ((task->tk_flags & RPC_TASK_NO_RETRANS_TIMEOUT)
 882                     && xprt_connected(xprt)
 883                     && req->rq_connect_cookie == xprt->connect_cookie) {
 884                         xprt->ops->set_retrans_timeout(task);
 885                         rpc_sleep_on(&xprt->pending, task, xprt_timer);
 886                         goto out_unlock;
 887                 }
 888         }
 889         if (!xprt->ops->reserve_xprt(xprt, task)) {
 890                 task->tk_status = -EAGAIN;
 891                 goto out_unlock;
 892         }
 893         ret = true;
 894 out_unlock:
 895         spin_unlock_bh(&xprt->transport_lock);
 896         return ret;
 897 }
 898
 899 void xprt_end_transmit(struct rpc_task *task)
 900 {
 901         xprt_release_write(task->tk_rqstp->rq_xprt, task);
 902 }
 903
 904 /**
 905  * xprt_transmit - send an RPC request on a transport
 906  * @task: controlling RPC task
 907  *
 908  * We have to copy the iovec because sendmsg fiddles with its contents.
 909  */
 910 void xprt_transmit(struct rpc_task *task)
 911 {
 912         struct rpc_rqst *req = task->tk_rqstp;
 913         struct rpc_xprt *xprt = req->rq_xprt;
 914         int status, numreqs;
 915
 916         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
 917
 918         if (!req->rq_reply_bytes_recvd) {
 919                 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
 920                         /*
 921                          * Add to the list only if we're expecting a reply
 922                          */
 923                         spin_lock_bh(&xprt->transport_lock);
 924                         /* Update the softirq receive buffer */
 925                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
 926                                         sizeof(req->rq_private_buf));
 927                         /* Add request to the receive list */
 928                         list_add_tail(&req->rq_list, &xprt->recv);
 929                         spin_unlock_bh(&xprt->transport_lock);
 930                         xprt_reset_majortimeo(req);
 931                         /* Turn off autodisconnect */
 932                         del_singleshot_timer_sync(&xprt->timer);
 933                 }
 934         } else if (!req->rq_bytes_sent)
 935                 return;
 936
 937         req->rq_xtime = ktime_get();
 938         status = xprt->ops->send_request(task);
 939         if (status != 0) {
 940                 task->tk_status = status;
 941                 return;
 942         }
 943
 944         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
 945         task->tk_flags |= RPC_TASK_SENT;
 946         spin_lock_bh(&xprt->transport_lock);
 947
 948         xprt->ops->set_retrans_timeout(task);
 949
 950         numreqs = atomic_read(&xprt->num_reqs);
 951         if (numreqs > xprt->stat.max_slots)
 952                 xprt->stat.max_slots = numreqs;
 953         xprt->stat.sends++;
 954         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
 955         xprt->stat.bklog_u += xprt->backlog.qlen;
 956         xprt->stat.sending_u += xprt->sending.qlen;
 957         xprt->stat.pending_u += xprt->pending.qlen;
 958
 959         /* Don't race with disconnect */
 960         if (!xprt_connected(xprt))
 961                 task->tk_status = -ENOTCONN;
 962         else {
 963                 /*
 964                  * Sleep on the pending queue since
 965                  * we're expecting a reply.
 966                  */
 967                 if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task))
 968                         rpc_sleep_on(&xprt->pending, task, xprt_timer);
 969                 req->rq_connect_cookie = xprt->connect_cookie;
 970         }
 971         spin_unlock_bh(&xprt->transport_lock);
 972 }
 973
 974 static void xprt_add_backlog(struct rpc_xprt *xprt, struct rpc_task *task)
 975 {
 976         set_bit(XPRT_CONGESTED, &xprt->state);
 977         rpc_sleep_on(&xprt->backlog, task, NULL);
 978 }
 979
 980 static void xprt_wake_up_backlog(struct rpc_xprt *xprt)
 981 {
 982         if (rpc_wake_up_next(&xprt->backlog) == NULL)
 983                 clear_bit(XPRT_CONGESTED, &xprt->state);
 984 }
 985
 986 static bool xprt_throttle_congested(struct rpc_xprt *xprt, struct rpc_task *task)
 987 {
 988         bool ret = false;
 989
 990         if (!test_bit(XPRT_CONGESTED, &xprt->state))
 991                 goto out;
 992         spin_lock(&xprt->reserve_lock);
 993         if (test_bit(XPRT_CONGESTED, &xprt->state)) {
 994                 rpc_sleep_on(&xprt->backlog, task, NULL);
 995                 ret = true;
 996         }
 997         spin_unlock(&xprt->reserve_lock);
 998 out:
 999         return ret;
1000 }
1001
1002 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt, gfp_t gfp_flags)
1003 {
1004         struct rpc_rqst *req = ERR_PTR(-EAGAIN);
1005
1006         if (!atomic_add_unless(&xprt->num_reqs, 1, xprt->max_reqs))
1007                 goto out;
1008         req = kzalloc(sizeof(struct rpc_rqst), gfp_flags);
1009         if (req != NULL)
1010                 goto out;
1011         atomic_dec(&xprt->num_reqs);
1012         req = ERR_PTR(-ENOMEM);
1013 out:
1014         return req;
1015 }
1016
1017 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1018 {
1019         if (atomic_add_unless(&xprt->num_reqs, -1, xprt->min_reqs)) {
1020                 kfree(req);
1021                 return true;
1022         }
1023         return false;
1024 }
1025
1026 void xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1027 {
1028         struct rpc_rqst *req;
1029
1030         spin_lock(&xprt->reserve_lock);
1031         if (!list_empty(&xprt->free)) {
1032                 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1033                 list_del(&req->rq_list);
1034                 goto out_init_req;
1035         }
1036         req = xprt_dynamic_alloc_slot(xprt, GFP_NOWAIT|__GFP_NOWARN);
1037         if (!IS_ERR(req))
1038                 goto out_init_req;
1039         switch (PTR_ERR(req)) {
1040         case -ENOMEM:
1041                 dprintk("RPC:       dynamic allocation of request slot "
1042                                 "failed! Retrying\n");
1043                 task->tk_status = -ENOMEM;
1044                 break;
1045         case -EAGAIN:
1046                 xprt_add_backlog(xprt, task);
1047                 dprintk("RPC:       waiting for request slot\n");
1048         default:
1049                 task->tk_status = -EAGAIN;
1050         }
1051         spin_unlock(&xprt->reserve_lock);
1052         return;
1053 out_init_req:
1054         task->tk_status = 0;
1055         task->tk_rqstp = req;
1056         xprt_request_init(task, xprt);
1057         spin_unlock(&xprt->reserve_lock);
1058 }
1059 EXPORT_SYMBOL_GPL(xprt_alloc_slot);
1060
1061 void xprt_lock_and_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1062 {
1063         /* Note: grabbing the xprt_lock_write() ensures that we throttle
1064          * new slot allocation if the transport is congested (i.e. when
1065          * reconnecting a stream transport or when out of socket write
1066          * buffer space).
1067          */
1068         if (xprt_lock_write(xprt, task)) {
1069                 xprt_alloc_slot(xprt, task);
1070                 xprt_release_write(xprt, task);
1071         }
1072 }
1073 EXPORT_SYMBOL_GPL(xprt_lock_and_alloc_slot);
1074
1075 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1076 {
1077         spin_lock(&xprt->reserve_lock);
1078         if (!xprt_dynamic_free_slot(xprt, req)) {
1079                 memset(req, 0, sizeof(*req));   /* mark unused */
1080                 list_add(&req->rq_list, &xprt->free);
1081         }
1082         xprt_wake_up_backlog(xprt);
1083         spin_unlock(&xprt->reserve_lock);
1084 }
1085
1086 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1087 {
1088         struct rpc_rqst *req;
1089         while (!list_empty(&xprt->free)) {
1090                 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1091                 list_del(&req->rq_list);
1092                 kfree(req);
1093         }
1094 }
1095
1096 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1097                 unsigned int num_prealloc,
1098                 unsigned int max_alloc)
1099 {
1100         struct rpc_xprt *xprt;
1101         struct rpc_rqst *req;
1102         int i;
1103
1104         xprt = kzalloc(size, GFP_KERNEL);
1105         if (xprt == NULL)
1106                 goto out;
1107
1108         xprt_init(xprt, net);
1109
1110         for (i = 0; i < num_prealloc; i++) {
1111                 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1112                 if (!req)
1113                         break;
1114                 list_add(&req->rq_list, &xprt->free);
1115         }
1116         if (i < num_prealloc)
1117                 goto out_free;
1118         if (max_alloc > num_prealloc)
1119                 xprt->max_reqs = max_alloc;
1120         else
1121                 xprt->max_reqs = num_prealloc;
1122         xprt->min_reqs = num_prealloc;
1123         atomic_set(&xprt->num_reqs, num_prealloc);
1124
1125         return xprt;
1126
1127 out_free:
1128         xprt_free(xprt);
1129 out:
1130         return NULL;
1131 }
1132 EXPORT_SYMBOL_GPL(xprt_alloc);
1133
1134 void xprt_free(struct rpc_xprt *xprt)
1135 {
1136         put_net(xprt->xprt_net);
1137         xprt_free_all_slots(xprt);
1138         kfree(xprt);
1139 }
1140 EXPORT_SYMBOL_GPL(xprt_free);
1141
1142 /**
1143  * xprt_reserve - allocate an RPC request slot
1144  * @task: RPC task requesting a slot allocation
1145  *
1146  * If the transport is marked as being congested, or if no more
1147  * slots are available, place the task on the transport's
1148  * backlog queue.
1149  */
1150 void xprt_reserve(struct rpc_task *task)
1151 {
1152         struct rpc_xprt *xprt;
1153
1154         task->tk_status = 0;
1155         if (task->tk_rqstp != NULL)
1156                 return;
1157
1158         task->tk_timeout = 0;
1159         task->tk_status = -EAGAIN;
1160         rcu_read_lock();
1161         xprt = rcu_dereference(task->tk_client->cl_xprt);
1162         if (!xprt_throttle_congested(xprt, task))
1163                 xprt->ops->alloc_slot(xprt, task);
1164         rcu_read_unlock();
1165 }
1166
1167 /**
1168  * xprt_retry_reserve - allocate an RPC request slot
1169  * @task: RPC task requesting a slot allocation
1170  *
1171  * If no more slots are available, place the task on the transport's
1172  * backlog queue.
1173  * Note that the only difference with xprt_reserve is that we now
1174  * ignore the value of the XPRT_CONGESTED flag.
1175  */
1176 void xprt_retry_reserve(struct rpc_task *task)
1177 {
1178         struct rpc_xprt *xprt;
1179
1180         task->tk_status = 0;
1181         if (task->tk_rqstp != NULL)
1182                 return;
1183
1184         task->tk_timeout = 0;
1185         task->tk_status = -EAGAIN;
1186         rcu_read_lock();
1187         xprt = rcu_dereference(task->tk_client->cl_xprt);
1188         xprt->ops->alloc_slot(xprt, task);
1189         rcu_read_unlock();
1190 }
1191
1192 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1193 {
1194         return (__force __be32)xprt->xid++;
1195 }
1196
1197 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1198 {
1199         xprt->xid = net_random();
1200 }
1201
1202 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1203 {
1204         struct rpc_rqst *req = task->tk_rqstp;
1205
1206         INIT_LIST_HEAD(&req->rq_list);
1207         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1208         req->rq_task    = task;
1209         req->rq_xprt    = xprt;
1210         req->rq_buffer  = NULL;
1211         req->rq_xid     = xprt_alloc_xid(xprt);
1212         req->rq_connect_cookie = xprt->connect_cookie - 1;
1213         req->rq_release_snd_buf = NULL;
1214         xprt_reset_majortimeo(req);
1215         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1216                         req, ntohl(req->rq_xid));
1217 }
1218
1219 /**
1220  * xprt_release - release an RPC request slot
1221  * @task: task which is finished with the slot
1222  *
1223  */
1224 void xprt_release(struct rpc_task *task)
1225 {
1226         struct rpc_xprt *xprt;
1227         struct rpc_rqst *req = task->tk_rqstp;
1228
1229         if (req == NULL) {
1230                 if (task->tk_client) {
1231                         rcu_read_lock();
1232                         xprt = rcu_dereference(task->tk_client->cl_xprt);
1233                         if (xprt->snd_task == task)
1234                                 xprt_release_write(xprt, task);
1235                         rcu_read_unlock();
1236                 }
1237                 return;
1238         }
1239
1240         xprt = req->rq_xprt;
1241         if (task->tk_ops->rpc_count_stats != NULL)
1242                 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
1243         else if (task->tk_client)
1244                 rpc_count_iostats(task, task->tk_client->cl_metrics);
1245         spin_lock_bh(&xprt->transport_lock);
1246         xprt->ops->release_xprt(xprt, task);
1247         if (xprt->ops->release_request)
1248                 xprt->ops->release_request(task);
1249         if (!list_empty(&req->rq_list))
1250                 list_del(&req->rq_list);
1251         xprt->last_used = jiffies;
1252         if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
1253                 mod_timer(&xprt->timer,
1254                                 xprt->last_used + xprt->idle_timeout);
1255         spin_unlock_bh(&xprt->transport_lock);
1256         if (req->rq_buffer)
1257                 xprt->ops->buf_free(req->rq_buffer);
1258         if (req->rq_cred != NULL)
1259                 put_rpccred(req->rq_cred);
1260         task->tk_rqstp = NULL;
1261         if (req->rq_release_snd_buf)
1262                 req->rq_release_snd_buf(req);
1263
1264         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1265         if (likely(!bc_prealloc(req)))
1266                 xprt_free_slot(xprt, req);
1267         else
1268                 xprt_free_bc_request(req);
1269 }
1270
1271 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1272 {
1273         atomic_set(&xprt->count, 1);
1274
1275         spin_lock_init(&xprt->transport_lock);
1276         spin_lock_init(&xprt->reserve_lock);
1277
1278         INIT_LIST_HEAD(&xprt->free);
1279         INIT_LIST_HEAD(&xprt->recv);
1280 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1281         spin_lock_init(&xprt->bc_pa_lock);
1282         INIT_LIST_HEAD(&xprt->bc_pa_list);
1283 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1284
1285         xprt->last_used = jiffies;
1286         xprt->cwnd = RPC_INITCWND;
1287         xprt->bind_index = 0;
1288
1289         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1290         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1291         rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending");
1292         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1293
1294         xprt_init_xid(xprt);
1295
1296         xprt->xprt_net = get_net(net);
1297 }
1298
1299 /**
1300  * xprt_create_transport - create an RPC transport
1301  * @args: rpc transport creation arguments
1302  *
1303  */
1304 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1305 {
1306         struct rpc_xprt *xprt;
1307         struct xprt_class *t;
1308
1309         spin_lock(&xprt_list_lock);
1310         list_for_each_entry(t, &xprt_list, list) {
1311                 if (t->ident == args->ident) {
1312                         spin_unlock(&xprt_list_lock);
1313                         goto found;
1314                 }
1315         }
1316         spin_unlock(&xprt_list_lock);
1317         printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1318         return ERR_PTR(-EIO);
1319
1320 found:
1321         xprt = t->setup(args);
1322         if (IS_ERR(xprt)) {
1323                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1324                                 -PTR_ERR(xprt));
1325                 goto out;
1326         }
1327         if (args->flags & XPRT_CREATE_NO_IDLE_TIMEOUT)
1328                 xprt->idle_timeout = 0;
1329         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1330         if (xprt_has_timer(xprt))
1331                 setup_timer(&xprt->timer, xprt_init_autodisconnect,
1332                             (unsigned long)xprt);
1333         else
1334                 init_timer(&xprt->timer);
1335
1336         if (strlen(args->servername) > RPC_MAXNETNAMELEN) {
1337                 xprt_destroy(xprt);
1338                 return ERR_PTR(-EINVAL);
1339         }
1340         xprt->servername = kstrdup(args->servername, GFP_KERNEL);
1341         if (xprt->servername == NULL) {
1342                 xprt_destroy(xprt);
1343                 return ERR_PTR(-ENOMEM);
1344         }
1345
1346         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1347                         xprt->max_reqs);
1348 out:
1349         return xprt;
1350 }
1351
1352 /**
1353  * xprt_destroy - destroy an RPC transport, killing off all requests.
1354  * @xprt: transport to destroy
1355  *
1356  */
1357 static void xprt_destroy(struct rpc_xprt *xprt)
1358 {
1359         dprintk("RPC:       destroying transport %p\n", xprt);
1360         del_timer_sync(&xprt->timer);
1361
1362         rpc_destroy_wait_queue(&xprt->binding);
1363         rpc_destroy_wait_queue(&xprt->pending);
1364         rpc_destroy_wait_queue(&xprt->sending);
1365         rpc_destroy_wait_queue(&xprt->backlog);
1366         cancel_work_sync(&xprt->task_cleanup);
1367         kfree(xprt->servername);
1368         /*
1369          * Tear down transport state and free the rpc_xprt
1370          */
1371         xprt->ops->destroy(xprt);
1372 }
1373
1374 /**
1375  * xprt_put - release a reference to an RPC transport.
1376  * @xprt: pointer to the transport
1377  *
1378  */
1379 void xprt_put(struct rpc_xprt *xprt)
1380 {
1381         if (atomic_dec_and_test(&xprt->count))
1382                 xprt_destroy(xprt);
1383 }
1384
1385 /**
1386  * xprt_get - return a reference to an RPC transport.
1387  * @xprt: pointer to the transport
1388  *
1389  */
1390 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1391 {
1392         if (atomic_inc_not_zero(&xprt->count))
1393                 return xprt;
1394         return NULL;
1395 }